Add EVO_Train bridge and VLA/RL training control plane

[pearlq12345]

Summary

This PR adds a RoboClaw-side cloud training control plane that talks to EVO_Train and exposes VLA/RL workflow contracts.

Key changes:

• Add an EVO_Train JSON/TCP bridge for cloud training requests
• Add cloud training APIs for plan, start, stop, current task, status, GPU SKUs, images, and runtime matching
• Add VLA/RL APIs for profiles, planning, artifact review, and deployability checks
• Add VLA/RL backend profiles with concrete contracts for RLinf, LeRobot, Dexbotic, and custom project backends
• Add policy registry support for ACT, Diffusion Policy, Pi0, GR00T, SmolVLA, and related policy families
• Add a minimal roboclaw_vla.rl adapter package with registry, launcher, evaluator, and adapter import surfaces
• Add a RoboClaw GRPO profile and RLinf-style Hydra config template chain
• Add TCP read/write timeout handling and clearer 502/503 error mapping
• Add robot-training skill guidance and VLA infra documentation

Note: roboclaw_vla.rl.launcher is intentionally a contract stub in this PR. It verifies the module/preflight surface and writes diagnostic artifacts, but full RLinf actor / rollout / env worker orchestration is outside the scope of this PR.

This pairs with the EVO_Train PR that adds AutoDL-backed VLA/RL workflow execution.

Validation

python -m pytest tests/test_vla_rl_routes.py tests/test_evo_train_routes.py tests/test_command_builder_train.py tests/test_policy_registry.py -q